Classification Matters: Divergent Estimates of Dementia Risk Factors in the Health and Retirement Study

Abstract Clinical dementia diagnosis requires substantial expertise and is complex and costly. Therefore, large population-based studies often lack gold-standard longitudinal diagnoses and instead use classification algorithms employing cognitive, functional, and health-related measures to identify dementia. We evaluated whether algorithm choice affects conclusions regarding dementia risk in the nationally representative Health and Retirement Study (HRS; 2010-2020). We constructed four algorithm-specific samples of participants age 70+ without dementia in 2010: Langa-Weir (n=6,498), Hudomiet (n=6,556), Expert (n=5,751), and Wu (n=5,935). Participants were followed until first dementia classification, death, or censoring. We estimated dementia incidence rates (IRs) and used adjusted Cox models to estimate hazard ratios (HRs) for known risk factors. Dementia IRs per 1000 person-years meaningfully varied across algorithms, from 26.5 (95% CI 24.5–28.5, Hudomiet) to 62.8 (59.9–65.8, Expert). Sociodemographic risk factor associations also varied substantially. HRs for education (< vs. ≥high school) ranged from 1.36 (1.19–1.54, Hudomiet) to 2.22 (1.98–2.50, Langa-Weir), and HRs non-Hispanic Black vs. non-Hispanic White ranged from 1.86 (1.53–2.27, Expert) to 2.74 (2.40–3.12, Langa-Weir). Differences were less pronounced for health-related risk factors. In the absence of gold standard longitudinal diagnosis in HRS, researchers should examine the robustness of findings across multiple algorithms.